xirt-torch 2.0.1

Retention-time prediction for linear and crosslinked peptides.
Modification-aware sequences. Shared recurrent encoders. PyTorch on CPU or CUDA.

Documentation · Quick start · Notebooks · Benchmarks · PyPI

What xiRT does

xiRT learns peptide embeddings and predicts continuous retention times or fraction labels. Crosslinked peptides share a recurrent encoder between their two sequence branches; linear peptides use a single input. Train a model for your experiment, use a saved checkpoint for prediction, or adapt a recurrent model with full fine-tuning, a frozen encoder, LoKr, or LoRA.

Install

Requires Python 3.12–3.14. The PyPI distribution is xirt-torch; the Python import and command remain xirt.

pip install xirt-torch

Generate the two configuration files, then edit their input columns and targets before training:

xirt -p xirt_params.yaml -s learning_params.yaml
xirt -i psms.csv -o results -x xirt_params.yaml -l learning_params.yaml

For a locked source installation, clone this repository and run uv sync --locked --extra cpu (or --extra cuda), then prefix the commands with uv run --no-sync. See installation and CLI for backend selection and optional tracking, TensorBoard, Lion, and ESM-C extras.

Learn with real peptide data

Two CPU notebooks use 2,048 observed peptide pairs:

From the source checkout:

uv run --no-sync marimo edit examples/train_predict.py --host 127.0.0.1
uv run --no-sync marimo edit examples/fine_tune.py --host 127.0.0.1

make examples executes both without a notebook server. Training uses up to 150 epochs with validation-based early stopping. The fine-tuning guide covers checkpoint compatibility, the Python API, and YAML settings.

Results and compatibility

The benchmark documentation links protocols, predictions, checkpoint provenance, and plots. PyTorch had lower mean held-out RMSE than TensorFlow on two of five evaluated cohorts, not all five. Full fine-tuning had the lowest mean RMSE among the four native fine-tuning strategies on all five cohorts under the fixed comparison protocol.

Upgrading from 1.x: TensorFlow is no longer an application dependency or supported backend, and xiRT 2.0 cannot load Keras .h5 models. See the migration guide. Historical artifacts remain in benchmark-tf-torch/<cohort>/results/.

Development

make sync
make check
make build

pixi run sync bootstraps the same locked uv environment. make help lists documentation, notebook, test and release commands. Benchmark code lives in xirt/data_pipelines/benchmarks/ and is excluded from the production wheel.

Configuration · Migration · Python API · Development

Citation

Giese, S.H., Sinn, L.R., Wegner, F. et al. Retention time prediction using neural networks increases identifications in crosslinking mass spectrometry. Nature Communications 12, 3237 (2021). DOI: 10.1038/s41467-021-23441-0

Contributors: Sven Giese and Ludwig Sinn.